EpiDistill

Geometric Distillation from Rectified Stereo: Leveraging Epipolar Cues for Monocular Depth

Jung-Hee Kim, Xiaoming Liu · Michigan State University / UNC Chapel Hill

Paper · Project page · Code

EpiDistill turns a frozen depth foundation model into a single-view metric depth model. A multi-view teacher learns correspondence through depth-guided epipolar attention; at inference the source views are replaced by learnable Rectified Stereo Tokens, so the cross-view pathway survives with one image.

Files

file	backbone	tensors
epidistill_unidepthv2.pt	UniDepthV2 (ViT-L/14)	401
epidistill_depthpro.pt	DepthPro	455

Each file is a plain PyTorch state_dict (no pickled objects beyond tensors, so it loads with weights_only=True).

These weights are a correction, not a depth network

They hold only EpiDistill's own parameters. The frozen backbone is not included and must be obtained separately:

• UniDepthV2 downloads automatically from lpiccinelli/unidepth-v2-vitl14.
• DepthPro needs depth_pro.pt from apple/ml-depth-pro.

Usage

from huggingface_hub import hf_hub_download
from epidistill import EpiDistillConfig, EpiDistillPredictor

path = hf_hub_download("kimjun84/EpiDistill", "epidistill_unidepthv2.pt")
predictor = EpiDistillPredictor.from_pretrained(
    path, config=EpiDistillConfig(backbone="unidepthv2")
)
depth, K = predictor(rgb)      # depth: (H, W) float32 metres, K: (3, 3)

The model code lives in the EpiDistill repository; install it first.

License

CC BY-NC 4.0 — non-commercial. The UniDepthV2 backbone these weights correct is itself CC BY-NC 4.0; DepthPro carries Apple's licence.

Citation

@inproceedings{kim2026epidistill,
  title     = {Geometric Distillation from Rectified Stereo:
               Leveraging Epipolar Cues for Monocular Depth},
  author    = {Kim, Jung-Hee and Liu, Xiaoming},
  booktitle = {ECCV},
  year      = {2026}
}